What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Keep causally related work for one request or agent turn in the same OpenTelemetry trace, propagate its context across boundaries, and represent orchestration, model inference, and tool execution as distinct spans. Use parent-child relationships for work that fits the operation’s actual nesting; use span links when important causal relationships do not fit a single-parent tree.
Start with the trace boundary
A trace is the record of work for one logical operation, such as handling an application request or running an agent turn. Its spans have individual SpanIds and share a TraceId. A span with a parent records that parent relationship and shares the parent’s TraceId. The root span should represent the real orchestration boundary—not simply whichever library happens to make the first model call.
As an Amazon Associate I earn from qualifying purchases.
For an agent workflow, that boundary is often the application’s invocation of the agent. For a simpler application with no agent orchestration, it may be the request handler or another operation that owns the model call. Instrument the actual boundaries in your application rather than imposing an identical tree on every framework.
Recommended Free Tools
Represent orchestration, inference, and tools separately
Each span should describe one operation. An agent span describes orchestration; an inference span describes the model request; and a tool span describes the tool’s execution. Keeping those operations distinct makes it possible to see whether time and errors belong to planning, a model request, dispatch, or tool work.
#1 Best Overall
Agent invocation
The OpenTelemetry GenAI agent convention recommends invoke_agent for agent invocation. It recommends an INTERNAL span kind for same-process agent work and a CLIENT kind for a remote invocation. These are evolving GenAI conventions, not universal behavior guaranteed by every SDK or framework.
The convention describes a plan span for planning work. If an LLM call generates the plan, that inference span is a child of the plan span. Tool or task operations arising from the plan are typically sibling operations under the agent invocation. Use this as a semantic guide: the recorded nesting should reflect the work your application actually performs.
Model inference
Record a model request as its own GenAI inference span. The GenAI client convention describes the logical operation as lasting until the response is fully received, or until the operation ends through error or cancellation; automatic retries belong to that logical operation. Give the span a useful operation name and include relevant GenAI attributes that the instrumentation actually provides.
Free tools Windows power users keep installed
One-click scans. No signup required.
Do not assume that a particular instrumentation library emits every GenAI attribute, or that a convention’s current names and requirements are stable. The official GenAI client and agent conventions are marked Development, so check the current convention and the documentation for your chosen SDK or framework before relying on specific fields.
Tool execution
A model request that asks for a tool and the tool’s actual execution are different operations. The GenAI client convention uses execute_tool for the tool operation and encourages application developers to manually instrument tool calls that automatic instrumentation does not cover.
Place execution under the active agent workflow or tool-dispatch operation when that matches the real call path. A subsequent model request is a later operation within the relevant agent invocation, not part of the tool’s execution span. If existing MCP instrumentation already represents a tool execution, avoid adding a second span for the same operation. Do not invent a tool-call identifier attribute; use one only when the current convention or framework documents it.
Propagate context across asynchronous and service boundaries
OpenTelemetry Context carries execution-scoped values across API boundaries and between logically associated units of work. In practice, the active trace context must follow work as it moves through the application; otherwise, an asynchronous task or downstream service may start a disconnected trace instead of appearing as part of the original request.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11- Establish context at the boundary. Ensure the incoming request or logical agent turn has an active trace context and a span representing the operation that owns the work.
- Carry or activate it across in-process work. Pass or activate context when execution crosses library, callback, or asynchronous boundaries, using the mechanisms supported by the language SDK and framework.
- Inject and extract at service boundaries. Configure OpenTelemetry propagators to carry context with outgoing calls and recover it in the receiving service. The W3C Trace Context propagator validates and carries supported
traceparentandtracestatevalues. - Check the resulting relationship. When propagation and instrumentation work together, the downstream operation appears in the same trace with the outgoing operation as its remote parent.
SpanContext is the identifying context used for span propagation; it includes the TraceId and SpanId and conforms to W3C Trace Context. Propagation preserves trace relationships across supported boundaries, but it does not by itself create useful spans for uninstrumented work.
Choose parent-child relationships or links based on causality
Parent-child nesting expresses a coherent operation with sub-operations. It is the natural choice when a model call, tool dispatch, or other unit of work is performed as part of an active operation and the tree reflects that execution.
Rank #4
A span has one parent, so a tree cannot always express every meaningful causal relationship. OpenTelemetry span links can point to related spans in the same or different traces. Use a link when an operation has causal predecessors that cannot be accurately represented by making one of them its parent. Do not use links as a substitute for ordinary nesting when the operation really is a child in the active workflow.
Keep conversation identity separate from trace identity
A conversation or session can outlive a single trace. The GenAI agent convention says to set gen_ai.conversation.id only when the instrumented library has the value readily available or the application supplies it through OpenTelemetry Context or a library-specific mechanism. Treat it as supplementary correlation where appropriate; it does not replace propagation of trace context.
Check instrumentation and backend coverage
Instrumentation coverage is not universal. Before adopting a library or trace backend, check which model providers and agent frameworks it instruments, whether application-owned tools need manual spans, and whether context survives the asynchronous and service boundaries in your architecture. Confirm that emitted names and attributes match the current conventions where that matters.
Best Value
- Used Book in Good Condition
Also verify that the backend lets you inspect parent-child and linked-span relationships, and decide how prompt, response, and tool data will be handled under your data-governance requirements. Do not assume every instrumentation captures message content or that capturing it is appropriate.
Amazon OpenSearch Service is one documented example of an AI observability backend built on OpenTelemetry and GenAI conventions; AWS describes hierarchical traces across orchestration, LLM calls, tool invocations, and retrieval. It is an optional backend example, not a requirement for instrumenting OpenTelemetry.
A practical shape for one agent turn
A trace might be organized conceptually like this, with nesting chosen to match the application’s actual execution:
- Agent invocation (
invoke_agent)- Plan operation
- Inference operation that generates the plan
- Tool dispatch or task operation
- Tool execution (
execute_tool)
- Tool execution (
- Later inference operation, such as a model request after tool output
- Plan operation
This is a guide to the meaning of spans, not a required tree for every agent implementation. If real execution differs, instrument that execution; if causal work cannot be represented correctly by the tree’s parent relationships, add links where appropriate.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
